AI agents running on Gemini 4 Argon and the Antigravity platform designed, simulated, and built a 12-DoF robot without manual CAD
Most robots start life as a stack of CAD files, a whiteboard full of joint diagrams, and an engineering team that hasn’t slept much. Google DeepMind’s latest one started as a sentence.
On October 31, 2026, DeepMind showed off a physical dancing robot that AI agents designed, simulated, and built from one plain-English prompt. Nobody on the team touched a manual CAD workflow.
From prompt to dance floor
The robot has 12 degrees of freedom, or DoF. A degree of freedom is one independent direction a machine can move. A robot with 12 of them has a dozen separate axes it can bend, twist, or rotate through, which is plenty to pull off something that passes for choreography.
Given the date of the demo, DeepMind gave the robot a Halloween theme.
AI agents handled the project end-to-end, moving from a text description to a design, through simulation, and on to a physical build.
Two pieces of DeepMind technology powered the effort. The first is Gemini 4 Argon, a model announced in late September 2026 and released broadly in mid-October 2026.
Gemini 4 Argon is built for complex reasoning that unfolds across many steps. It has a particular focus on robotics simulation and on generating accurate representations from video inputs.
The second piece is Antigravity, a development platform DeepMind launched in 2025. Antigravity is designed for autonomous coding and for orchestrating work across multiple systems.
Together, they reflect what the research describes as a move toward agent-first environments in robotics. In that kind of setup, AI agents are the primary operators, and humans step back into a supervisory role.
The Boston Dynamics connection
The project was led by Aaron Saunders, who previously served as VP of Engineering at Boston Dynamics before joining DeepMind.
The dancing robot also extends a line of robotics work at DeepMind. In September 2025, the lab introduced RoboBallet, a system aimed at improving how collaborative robots work together in manufacturing settings.
RoboBallet focused on coordination: getting multiple machines to share a workspace efficiently. The October 2026 demonstration pushes further upstream, into the question of how robots get designed in the first place.
Why a dancing robot is more serious than it looks
Hardware design has traditionally been one of the most labor-intensive parts of building a robot. Every joint, motor placement, and structural element typically requires careful manual modeling before anything gets fabricated. Removing manual CAD from that loop, even for a relatively simple 12-DoF machine, attacks a major bottleneck.
For the wider tech sector, the demonstration is another data point in the race over embodied AI. That term refers to AI systems that act in the physical world, not just on a screen.
Gemini 4 Argon has only been broadly available since mid-October 2026, which means outside developers are just beginning to test what it can do.
